Modeling of an unmanned aerial vehicle and autonomous control based on deep reinforcement learning

Research interest in Unmanned Aerial Vehicles (UAV) has increased rapidly due to their potential uses for a wide variety of applications. UAVs offer a lot of promise with their ability to perform dangerous and continuous tasks in remote and risky environments, thanks to the ability of UAVs such as, perform flight maneuvers, ball throwing, and catching, and order control of Unmanned Ground Vehicles, the UAV's ability to perform samples by acting as a flying sensor in critical agricultural areas and to undertake many similar tasks. In recent years, great progress in information technology and artificial intelligence has significantly affected the increasingly widespread intelligent autonomous systems. This kind of development has also influenced Unmanned Aerial Vehicles, which consist of fully autonomous, semi-autonomous, and remotely controlled flying vehicles. The primary objective of the studies in this field is to create an autonomous system that can make its own decisions such as planning its own path, maneuvering, and returning to the duty place according to the current situation and conditions without any human control. The main task expected from an Autonomous Unmanned Aerial Vehicle is to avoid collisions without human intervention and to lead to desired targets along an efficient path. Therefore, it plays a very critical role for a UAV to perform autonomous missions in extreme environments. This thesis, it is aimed to create an autonomous UAV path planning algorithm using a deep reinforcement learning approach. The main purpose of this thesis is to reach the determined target by overcoming certain three-dimensional static obstacles. For this, it is expected that the UAV will train itself and reach the determined target by using the trained network structure. In this thesis, in order for the UAV to reach its determined target, it has to create its own path by passing from the right or left of the obstacles. For this purpose, a Deep Determining Policy Gradient (DDPG) with a continuous action area was designed in this thesis. At the same time, the control of the UAV is communicated to the UAV via the MAVROS communication protocol. This communication protocol has been redesigned specifically for this thesis. A reward function specific to this thesis was developed to minimize the distance between the UAV and the destination and penalize collisions. In this way, it was ensured that the UAV avoided collisions and reached the desired target by the shortest route.